Primary use stated or inferred cautiously from official documentation.
N
NVIDIA / nvidia/Nemotron-3.5-Rerank-8B-BF16
Nemotron 3.5 Rerank 8B BF16
NVIDIA reranker with 8.38B parameters, released on October 20, 2026. It scores each text-query and document pair; documents may contain text, a page image or both. It complements initial retrieval by reranking candidates; it does not generate responses or embeddings. It combines a Pixtral vision encoder with a bidirectional Ministral3 tower.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | nvidia/Nemotron-3.5-Rerank-8B-BF16 |
|---|---|
| Model type | Reranking · Vision and documents · Multimodal |
| Access model | Open weights |
| License | OpenMDW 1.1 |
| Deployment | Self-hosted deployment with NVIDIA GPUs |
| Release | 20 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image |
| Exit | Scores |
| Context window | 32,768 tokens shared between text and image |
| maximum output | One relevance logit per query-document pair |
| Reasoning | Information retrieval, not conversational generation |
| Published tools | Candidate reranking |
| Structured Outings | Uncalibrated logits; higher values indicate greater relevance |
| Batch processing | Batch inference on self-hosted infrastructure |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · Transformers · Sentence Transformers · vLLM |
Semantic search and RAG over reports, tables, charts and infographics. NVIDIA reports text evaluation in 34 languages and visual evaluation in 18; this coverage is not a universal guarantee. Use the exact checkpoint identifier and official query and document prefixes. The examples support batch inference on self-hosted infrastructure.
The documented limit is 32,768 combined tokens, although the internal configuration declares more positions. Images default to a longest edge of 1,120 pixels, expandable to 1,540. Review any trust_remote_code before execution. The Transformers examples with Flash Attention 2 require version 5.20.0; vLLM 0.32.0 uses a separate environment with Transformers below 5.20.0. Inferama has not run inference.
03 / CAPABILITIES
What it can do
Orientation
Semantic search and RAG over reports, tables, charts and infographics. NVIDIA reports text evaluation in 34 languages and visual evaluation in 18; this coverage is not a universal guarantee. Use the exact checkpoint identifier and official query and document prefixes. The examples support batch inference on self-hosted infrastructure.
Context
32,768 tokens shared between text and image · One relevance logit per query-document pair
Tools and integration
Candidate reranking
Access
Hugging Face · downloadable checkpoint
Input modalities
TextImage
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
Public, ungated weights under OpenMDW 1.1, with commercial use subject to its terms and preservation of notices. The Ministral bases retain their Apache 2.0 terms. No hosted commercial API, identifier or price has been verified: numeric cost fields remain empty. Downloading does not eliminate GPU, memory and operating costs.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
The official model card mentions ViDoRe V3 and MIRACL-VISION with NDCG@10, but publishes no verifiable numerical scores for this checkpoint. We do not transfer predecessor results or invent ranking positions. Inferama evaluation is pending. For reproduction, retain revision, dataset, languages, candidates, resolution, dimensions, normalization, hardware and latency; compare retrieval and reranking as separate stages.
i ViDoRe V3 and MIRACL-VISION documented; no verified numerical scores.
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
nvidia/Nemotron-3.5-Rerank-8B-BF16
Release date stated in the official model card; do not confuse it with repository creation or modification.
07 / SOURCES